____________________________________________________________________________________________________ GEOGRAPHIC INFORMATION SYSTEMS AND WATER RESOURCES IV AWRA SPRING SPECIALTY CONFERENCE

advertisement
____________________________________________________________________________________________________
GEOGRAPHIC INFORMATION SYSTEMS AND WATER RESOURCES IV
AWRA SPRING SPECIALTY CONFERENCE
Houston, Texas
May 8-10, 2006
Copyright © 2006, AWRA
____________________________________________________________________________________________________
EXTRACTING AND CLASSIFYING BARE SOIL EROSION RISK AREAS IN A URBAN BASIN USING OBJECTORIENTED TECHNOLOGIES, HIGH RESOLUTION IMAGERY AND ELEVATION DATA
R. A. A. Nobrega, C. G. O’Hara, V. Vijayaraj, G. Olson, S. Kim
J. A. Quintanilha* and M. T. L. Barros*.
ABSTRACT: Investigating urban basin characteristics is of extreme importance for planning and managing the environment.
In areas of moderate to rugged terrain, uncontrolled terrace building and soil removal are considered harmful to drainage
basins and in-stream water quality. Identifying and classifying areas with potential for erosive soil can help prevent
environmental damages. However, detecting and measuring these activities within irregular settlements on the outskirts of
large cities are difficult tasks, especially in developing countries. This paper presents a study area near the urbanized extent of
Sao Paulo City-Brazil, characterized by recent and unplanned occupation, where dense urbanization is found intermixed with
preserved nature areas in the upstream part of the basin. In this tropical location, areas are normally affected by storms and
seasonal high-water events. The impact of soil erosion on in-stream flow has enlarged areas subject to flood inundation. The
methodology presented employs high resolution imagery to depict irregular settlements, vegetation coverage, and bare soils.
Although easily visualized on images, classifying bare soils according to the potential risk for erosion and producing relative
risk maps is problematic. The large amount of data requires automated procedures of image classification. However, the high
internal variance of these images causes incompatibility with most traditional classifiers. To overcome this shortcoming of
traditional approaches, image classification for this effort utilizes object-oriented technologies. First, the image is segmented
and the objects are detected through spectral, spatial and topological characteristics. Relationships among objects are used for
final fuzzy classification. The accuracy of 81.3% shows the potential of the methodology. Later, pre-classified objects are
analyzed considering information from a digital terrain model. The classification provides information about the relative
susceptibility to erosion for areas with bare soils.
KEY TERMS: object-based classification, high resolution image, erosion, urban basin, water.
INTRODUCTION
Recent developments on digital sensors and have stimulated new methodologies for land cover/land use classification. The
newer generation of satellite sensors, increased in radiometry as well as spectral and spatial resolutions, should be considered
as a new paradigm to be explored. Concerning the spatial resolution, Neubert & Meinel (2005) reports the improvement on
urban object discriminations and the reduction of spectral mixing of pixels. Tso & Mather (2001) and Repaka et al (2004)
agree that the efficiency of traditional per-pixel classification is reduced for high resolution images. However, investigations
on classification from high resolution images have shown significant increases in accuracy when object-based approaches
were used. The growing need for precise geoinformation for the GIS community generates the necessity for new research and
new applications.
The object-based classification proposed in this methodology aims the precise extraction of bare soils from a study area
located in an urban region of Sao Paulo city – Brazil. The region is characterized by recent and unplanned occupation, where
dense urbanization is found intermixed with preserved natural areas in the upstream part of the basin. Some tropical storms
and seasonal high-water events that affect this area, composed by moderate to rugged terrain, have caused impact of soil
erosion on in-stream flow and have enlarged areas subject to flood inundation.
The present work reports a methodology that extracts polygons of bare soil areas with high accuracy through segmentation
and classification according to the slope map by using GIS analysis. The goal is to create a methodology that provides precise
geoinformation to support hydrologic and land management works.
____________________________________________________________________________________
*
Respectively, Research Associate, Associate Professor, Research Associate, Research Associate,
Research Associate, Associate Professor, Associate Professor. GeoResources Institute, Mississippi State
University, Mississippi State-MS 39759, Phone: (662) 325-5114, Fax: (662) 325-7692. Escola
Politecnica da Universidade de Sao Paulo, Sao Paulo, Brazil. E-Mail: rodrigo@gri.msstate.edu
2006 AWRA Spring Specialty Conference
May 8-10, 2006
Data Sets and Study Area
A multi-spectral Ikonos image, recorded in 2002 over the north region of town was used for this research. The study area
covers 2.75 x 1.5 Km located in the upstream part of the Cabucu de Baixo River basin, between the latitudes 23026’26”SS
and 23027’16”S and longitudes 46040’14”W and 46041’28”W.
Figure 1. General visualization of north region of Sao Paulo (extracted from Google Earth) and the detailed study area.
Originally, the imagery was composed of NIR, Red, Green and Blue bands with 11 bits, orthorectified and resampled to 1
meter spatial resolution. However, the Principal Components (PC) were previously calculated and inserted on the project on
behalf of providing a large range of separability among objects in the scene.
The methodology also required a slope map generated from a precise digital terrain model. This data was prepared in 2004
covering the full basin and previously used on the Decision Support System. In this research, the slop map was subset
according to the limits of study area The Definiens eCognition was utilized to extract bare soil polygons and slope analysis
was performed using ESRI ArcGIS.
OBJECT-BASED CLASSIFICATION OF THE BARE SOIL AREAS
Object-based classification uses a different approach to assign classes to the objects. After creating the objects a large amount
of characteristics can be extracted, regarding spectral, geometric and topological domains. The classification rules can be
constructed supported by the object characteristics.
Segmentation
Due to the importance of properties of the objects, the segmentation required extensive analyses to fit the desired urban
objects relative to the segments. Variations of scale parameter, bands, degree of shape and compactness homogeneity were
tested. The best result was a combination of all original bands, scale parameter of 30, shape homogeneity of 0.9 and
compactness of 0.1.
An important observation can be considered for the shape homogeneity criterion. The larger the value, the lower the
consideration for color on the generation of the segments. It was necessary to detect small ceramic roofs.
Customized Features
Three customized features were implemented and used to classify the land cover. The Normalized Difference Vegetation
Index (NDVI) was employed to detect vegetation covered areas. According to Jensen (1996), NDVI has been used
extensively to measure vegetation amount on a worldwide basis. Originally developed in 1973, the NDVI computes the
2006 AWRA Spring Specialty Conference
May 8-10, 2006
normalized difference of brightness values from NIR and Red rationing and its use is disseminated today. Indeed, the
vegetation areas have higher values than others.
 µNIR − µRED 
NDVI = 

 µNIR + RED 
The second customized feature was created to point out shadows and other extremely dark areas in the remaining image. The
Shadow General Indicator uses a difference of brightness values between Blue band and the PC-1. Analyzing these two
images independently, the difference of brightness values for shadow and other similar objects are low in comparison with
the remaining objects of the image. However in different ranges, the subtraction of PC-1 values from Blue ones produces a
new map where the values are negatives. For the result image, the bright areas indicate the shadow objects that will be
detected and masked out latter.
Shadow _ General _ Indicator = µBlue − µPC1
Another feature was customized by using the combination of PC-3 and Blue toward the detection of bare soil areas. Despite
the PC-3 could discriminate the bare soil objects well, the excessive heterogeneity of the unplanned settlement makes more
difficult to analyse where the bare soil areas are mixed with the buildings and the streets. To improve the power of
discrimination, the Blue band was introduced because due to the low brightness for bare soil areas as well as the low variance
in comparison with other bands. Also, the equation uses a large and negative value on denominator just to produce small
values.
 µPC 3 * µBlue 
Bare _ General _ Indicator = 

 − 100000 
Classification
Three main classes were previously defined: Vegetation General, Shadow General and Bare General. However, the extraction
of bare soil areas began from the classification of vegetation areas by using the NDVI feature. Objects with NDVI value that
are higher then the threshold of 0.2 were considered as vegetation. Likewise, the shadow areas were detected for the rest of
the scene by using the Shadow General Indicator feature, with threshold of -425. These values were introduced to each class
definition to compose the fuzzy-soft scheme. Each object was classified according to the degree of pertinence of the class.
These two initial steps were necessary due to the complexity of the scene. Vegetation areas as well as shadow areas were
masked out from the remaining classification scheme to increase of detection of bare soil objects.
Figure 2. Combination of Red, Green and Blue bands of Ikonos image and the respective final classification.
Concerning the detection of the bare soil areas, two new classes were introduced -Ceramic Roof and Bare Soil– as childclasses of the Bare General class. The spectral patterns of ceramic roofs and bare soil areas are very similar. The introduction
of geometric and topological properties increased the power of the classification. Basically, the bare general objects were
analyzed according to the area, rectangular fitting and relative border to the Shadow General class. Objects in accordance
2006 AWRA Spring Specialty Conference
May 8-10, 2006
with the three conditions had higher potential to be classified as Ceramic Roofs. Otherwise, they were classified as Bare
Soils.
Table 1. Relative areas per class
CLASS
AREA (m2)
Vegetation General
1,982,186
Urbanized
827,189
Bare Soil
175,224
Shadow General
115,080
Ceramic Roof
19,046
TOTAL AREA
3,118,725
%
63.56
26.52
5.62
3.69
0.61
100
Relative Areas
Vegetation General
Shadow General
Ceramic Roof
Baren Soil
Urbanized
Figure 3. Relative area per class.
Certain steps were considered to validate the methodology regarding the accuracy of automatic procedure for bare soil
detection. First, all areas of bare soil compatible with the visualization scale of 1:2000 were digitalized on screen using an
appropriated combination of bands Red, Green and Blue. The digitized polygons were considered error-free. Then, a
comparison between these two sets of bare soil polygons was done. The accuracy of the automatic approach was 81.3 %.
Omission error of 18.7 % was partially generated by the cartographic generalization due to the resolution of the image. The
commission error of 28.7 % was generated due to very high similarity between the proprieties of ceramic roofs and some
bare soil polygons, especially inside the urbanized areas.
Table 2. Analysis of the of the classification for bare soil areas
METHOD
AREA (m2)
Manual (visual interpretation)
159,254.4
Automatic (object-based classification)
175,224.0
Correct classification (area of intersection)
129,479.4
200000.00
160000.00
area m2
Automatic
120000.00
Manual
Classified Correct
80000.00
40000.00
0.00
Figure 4. Comparison between manual and automatic results for the classification of bare soil areas
2006 AWRA Spring Specialty Conference
May 8-10, 2006
SLOPE CLASSIFICATION
After the extraction of the bare soil areas, the polygons were submitted to a new classification, according to the slope map.
This processing required the conversion of the polygons to a raster data to compute the intersection of both slope map and
bare soil images. During this processing, some cautions must be taken to preserve the compatibility between the spatial
resolutions of both sources. The less the pixel size, the precise the measuring of the areas over the new classified bare soil
image. The spatial resolution of the raster elevation data was one meter.
A new raster data was generated including 5 classes of information. The majority of bare soil polygons are located on areas
with low degree of slope, almost 86 %, represented in yellow on Figure 5. Almost 13% are composed by areas moderated to
rugged terrain. Only less than 1% is composed by very high slope, but their potential to surface erosion can produce
significant environmental damages. Figure 6 illustrates a superimposition between the digital terrain model, the orthophoto
and the detected bare soil polygons.
Figure 5. Superimposition of classified areas and slope map. Fragmentation of bare soil polygons according to the slope map.
Object_ID
0
1
2
3
4
Table 3. Classification of new bare soil polygons according to the slope map
Slope %
Area (m2) relative area % potential for superficial erosion
0 – 10
87,360
49.86
very low
10 – 25
63,408
36.19
low
25 – 50
18,478
10.54
medium
50 – 100
4,569
2.61
high
100 or more
1,409
0.80
very high
However some of these areas do not look very sensitive to surface erosion, most of them are located in flooding regions, low
and flat land alongside of watercourse. To do a decisive analysis of the potential for surface erosion the geological
characteristics of these areas must be considered.
Figure 6. Three-dimensional visualization of extracted bare soil polygons over the Ikonos orthorectified image.
2006 AWRA Spring Specialty Conference
May 8-10, 2006
DISCUSSION AND CONCLUSION
However the first experimentation of object-based classification to extract specific information of an unplanned settlement
from high resolution satellite imagery, the results show the high potential of this technology. In comparison to visual
interpretation, the accuracy of 81,3 % reports how precise the classification process can be, considering the extreme
heterogeneity of the spectral and spatial characteristics of the land cover.
Otherwise, nowadays, remote sensing and GIS are more integrated than never. The recent developments interfacing remote
sensing and GIS have obtained positive results, especially for high resolution images. In accordance to the set of these new
geo-applications, this paper presents a methodology to support information of the amount and location of bare soil areas. The
accuracy of the processing shows that the final mapping can provide good technical information to urban land/water
management. Based on the promising results, futures studies could be developed.
ACKNOLEDGMENTS
The authors would like to acknowledge the GeoResources Institute of Mississippi State University, the Escola Politecnica of
Universidade de Sao Paulo and the CAPES for the financial support.
REFERENCES
Baatz, M,; Bens, U.; Deghani, S.; Heynen, M.; Holtje, A.; Hofmann, P.; Lingenfelder, I.; Mimler, M.; Solhbahc, M.; Weber,
M.; Wilhauck, G. 2004. eCognition User Guide. Definiens Imaging GmbH, Munich, Germany.
Baatz, M.; Schape, A. 2001. Multiresolution Segmentation: an Optimization Approach for High Quality Multi-Scale Image
Segmentation. In: STROBL, J.; BLASCHKE, T., GRIESEBNER, G. Angewandte Geographische
Informatiosverarbeitung XII, Wichmann Verlag. Heidelberg. P. 12-23.
Blaschke, T.; Strobl, J. 2001. What’s Wrong With Pixels? Some Recent Developments Interfacing Remote Sensing and GIS.
In GIS-Zeitschrift fur Geoinformationsysteme, Helt 6, p 12-17.
Barros, M. T. L., 2004. Gerenciamento Integrado de Bacias Hidrograficas em Areas Urbanas, CNPq-EUSP, Sao Paulo.
Freauman, E.; Wolf, E. 2005. Segmentation of Very High Spatial Resolution Satellite Images in Urban Areas for SegmentsBased Classification. Proceedings for 3rd International Symposium Remote Sensing and Data Fusion Over Urban Areas.
Tempe, Arizona.
Jacobsen, K., 2001. New Developments in Digital Elevation Modeling. Geoinformatics, June 2001, pp18-21
Jensen, J. 1996. Introductory Digital Image Processing, A Remote sensing Perspective. 2nd Ed. New Jersey: Prentice-Hall.
Kresser, F. P.; Steinnocher, K.; Franzen, M. 2005. Object-Oriented Classification of Orthophotos to Support Update of
Spatial Databases. Proceedings for IGARS 2005 Symposium, Seoul, South Korea.
Neubert, M.; Meinel, G. 2005. Analise dos dados do satelite Ikonos baseada em segmentacao – utilizacao do software
eCognition para diferentes areas-teste. P. 115-124 In Blaschke T.; Kux, H. Sensoriamento Remoto e SIG: avancos,
novos sistemas sensores e metodos inovadores. Ed. Oficina de Textos, Sao Paulo.
Nobrega, R. A. A.; Quintanilha, J.A.; Barros, M. T. L. Emprego e Tecnicas Digitais para a Concepcao da Base Cartografica
Precisa da Bacia Hidrografica do Rio Cabucu de Baixo. Revista Brasileira de Recursos Hidricos, no prelo.
Repaka, S.; Truax, D.; Kolstad, E.; O’Hara, C. G. 2004. Comparing Spectral and Object-Based approaches for Classification
and Transportation Feature Extraction from High Resolution Multispectral Imagery. Proceeding for ASPRS 2004
Annual Conference.
Tso, B.; Mather, P. M. 2001. Classification Methods for Remotely Sensed Data. Taylor and Francis, London.
Download